91亚色 study highlights how deep-network models take potentially dangerous 鈥榮hortcuts鈥� in solving complex recognition tasks

TORONTO, Sept. 16, 2022 鈥�  Deep convolutional neural networks (DCNNs) don鈥檛 see objects the way humans do 鈥� using configural shape perception 鈥� and that could be dangerous in real-world AI applications, says , co-author of a 91亚色 study published today.

Published in the Cell Press journal , is a collaborative study by Elder, who holds the 91亚色 Research Chair in Human and Computer Vision and is Co-Director of 91亚色鈥檚 Centre for AI & Society, and Assistant Psychology Professor Nicholas Baker at Loyola College in Chicago, a former postdoctoral fellow at 91亚色.

The study employed novel visual stimuli called 鈥�Frankensteins鈥� to explore how the human brain and DCNNs process holistic, configural object properties.

鈥淔rankensteins are simply objects that have been taken apart and put back together the wrong way around,鈥� says Elder.  鈥淎s a result, they have all the right local features, but in the wrong places.鈥� 

The investigators found that while the human visual system is confused by Frankensteins, DCNNs are not 鈥� revealing an insensitivity to configural object properties.

鈥淥ur results explain why deep AI models fail under certain conditions and point to the need to consider tasks beyond object recognition in order to understand visual processing in the brain,鈥� Elder says. 鈥淭hese deep models tend to take 鈥榮hortcuts鈥� when solving complex recognition tasks. While these shortcuts may work in many cases, they can be dangerous in some of the real-world AI applications we are currently working on with our industry and government partners,鈥� Elder points out.

One such application is traffic video safety systems: 鈥淭he objects in a busy traffic scene 鈥� the vehicles, bicycles and pedestrians 鈥� obstruct each other and arrive at the eye of a driver as a jumble of disconnected fragments,鈥� explains Elder. 鈥淭he brain needs to correctly group those fragments to identify the correct categories and locations of the objects. An AI system for traffic safety monitoring that is only able to perceive the fragments individually will fail at this task, potentially misunderstanding risks to vulnerable road users.鈥�

According to the researchers, modifications to training and architecture aimed at making networks more brain-like did not lead to configural processing, and none of the networks were able to accurately predict trial-by-trial human object judgements. 鈥淲e speculate that to match human configural sensitivity, networks must be trained to solve a broader range of object tasks beyond category recognition,鈥� notes Elder.

91亚色 is a modern, multi-campus, urban university located in Toronto, Ontario. Backed by a diverse group of students, faculty, staff, alumni and partners, we bring a uniquely global perspective to help solve societal challenges, drive positive change and prepare our students for success. 91亚色's fully bilingual Glendon Campus is home to Southern Ontario's Centre of Excellence for French Language and Bilingual Postsecondary Education. 91亚色鈥檚 campuses in Costa Rica and India offer students exceptional transnational learning opportunities and innovative programs. Together, we can make things right for our communities, our planet, and our future. 

Media Contact:

Gloria Suhasini, 91亚色 Media Relations, 647.463.4354, suhasini@yorku.ca

 [KL1]Tried to eliminate one of the many 鈥渢o鈥� in this sentence

The post Study: Even smartest AI models don鈥檛 match human visual processing appeared first on News@91亚色.